An integrated scheme of arbitrarily shaped segmentation and motion estimation
Ryoichi Kawada, Atsushi Koike, Shuichi Matsumoto · Systems and Computers in Japan · 2000
In a rectangular-block division-based coding method, there is a problem that, in the blocks where different regions coexist, the coding efficiency decreases. In the segmentation-based coding method, which is promising as a way to avoid this problem, the objective of motion estimation and segmentation is to minimize the total generated entropy, that is, the sum of prediction error entropy, shape entropy, and motion vector entropy. The conventional motion estimation and segmentation methods based on processing such as split and merge, from this viewpoint, are not sufficiently optimized. Thus, in this paper, with the objective of optimizing segmentation-based motion-compensated prediction, the authors propose a scheme where motion estimation and segmentation are performed at the same time via dynamic programming. In the proposed scheme, first, for each motion vector and for each pattern of possible segment shapes, its generated entropy is estimated beforehand. Next, dynamic programming is applied to determine the segment for each motion vector that minimizes the total generated entropy. Computer simulation experiments using ordinary test video sequences show that this scheme outperforms the block matching method and conventional segmentation and motion estimation methods. The authors hope that this study will help to establish a general way to apply a segmentation-based coding method, where it has been pointed out that the efficiency is highly dependent on the kinds of pictures involved. © 2000 Scripta Technica, Syst Comp Jpn, 31(13): 19–30, 2000